A small language model that thinks through a cognitive architecture's numeric state. ⚠️ Needs the Sentiance runtime to run — the adapter alone is just a Qwen2.5-0.5B fine-tune; the hybrid behaviour comes from feeding the live state vector m_t through the trained state encoder. Author: Dr. Sanjay Anbu.

A small language model conditioned on a cognitive-architecture state vector. It is the "fused mind" of the Sentiance project (Path B / ADR 0005): a Qwen2.5-0.5B base with a LoRA adapter and a trained state encoder that turns the mind's whole-cycle state vector m_t into soft-prefix tokens, so generation is causally conditioned on the numeric state — not on state described in words.

Honest stance (non-negotiable). This is a functional artifact. m_t is a vector of functional variables (a valence, a drive level, a bond strength) named for the roles they play in the architecture. Conditioning a transformer on them buys integration and end-to-end learnability — not phenomenal experience. No claim of consciousness or sentience is made or implied. See ADR 0002.

What it is

Base: Qwen/Qwen2.5-0.5B-Instruct (Apache-2.0).
Adapter: LoRA (r=16, α=32) on the attention + MLP projections.
State encoder: a small MLP mapping the 41-dim m_t → n_prefix (default 16) soft-prefix embeddings, trained jointly with the LoRA on the next-token loss.
m_t encodes one tick of the cognitive cycle: valence/arousal/mood, a 14-way emotion one-hot, four drives, the attention source, goal presence, and 13 per-faculty signals (frustration, longing, empathy, grief, curiosity, anticipation…). See sentiance/mind/state_vector.py.
Trained state-blind: the felt state is removed from the prompt so m_t is the model's only state channel. (This is essential — see Results.)

Intended use

Research into cognitive architectures + LLMs, affective/state-conditioned generation, and neuro-symbolic integration. It is the inner voice of a Sentiance Mind: SENTIANCE_COGNITION_BACKEND=fused.

Not a general chat model, an assistant, or a source of factual answers.

How to run

This is not a plain AutoModelForCausalLM — the adapter alone is just a Qwen fine-tune. The fused behavior needs the Sentiance runtime to compute m_t each tick and inject it through the state encoder:

bash

pip install -e ".[finetune]"     # from the Sentiance repo
# place the model dir at models/sentiance-fused (adapter + state_encoder.pt + fused_config.json)
SENTIANCE_COGNITION_BACKEND=fused python -m sentiance chat

Results — does `m_t` actually steer it?

Measured with scripts/eval_fused.py (deterministic ablation over a 12-state battery spanning valence −0.80…+0.85; prompt held identical, only the vector changed) on a 12-state battery spanning valence −0.80…+0.85. 234 blended training examples.

The effect is real but data-scale-limited — report it as a distribution, not a single number. Across training seeds (3-seed sweeps):

conditioning	per-seed `r(valence, ΔAffect)`	mean ± std	strong (r≥0.5, p≤0.05)	mean `KL(real‖zero)`
prefix (soft tokens)	+0.82, +0.74, −0.46	+0.37 ± 0.59	2/3	0.004
FiLM (per-layer γ/β)	+0.83, −0.26, +0.63	+0.40 ± 0.47	2/3	0.10

The best seeds are strong and significant (r ≈ 0.8, permutation p ≈ 0.001–0.003, dose-response slope > 0), but ~1 seed in 3 fails and the variance is large — so the honest summary is directional but noisy, limited by the small dataset, not robust.

Two controls both give r ≈ 0: a shuffled m_t (structure destroyed), and the state-in-prompt model (state left in the words, so the vector is redundant and ignored). So the core claim holds as an ablation: state-as-vector conditions the model; state-as-text does not.

prefix vs FiLM (a negative result worth reporting): injecting m_t deep (FiLM, into every layer) makes its influence on the distribution ~25× larger (KL 0.10 vs 0.004) but does not improve seed-to-seed reliability — both are noisy at this data scale. The bottleneck is data, not conditioning depth.

Regenerate the numbers for any checkpoint with scripts/eval_fused.py (single) or scripts/robustness_fused.py (across seeds).

Limitations

Directional but noisy. Mean congruence r ≈ 0.4 across seeds with large std; ~1 seed in 3 does not learn the mapping. Not yet seed-robust — needs more data.
Small & blended. ~234 self-generated examples across several characters; the voice is repetitive.
0.5B base, English only, short first-person thoughts — not general text.
Self-generated data. Traces come from Sentiance's own (partly rule-based) cognitive cycle, so the model learns that system's regularities, not the world's.
Functional only. Nothing here evidences subjective experience.

Training data & reproducibility

Self-labeled traces exported from Sentiance runs (society / live / chat), deduplicated (incl. near-echo filtering), prepared state-blind. The full pipeline (collect → prepare → train → eval) and the ablation control are documented in the repo README and reproducible on a 6 GB laptop GPU.

License & attribution

This adapter + encoder: MIT (as the Sentiance repo).
Base model Qwen/Qwen2.5-0.5B-Instruct: Apache-2.0 — retain its notice; this is a derivative. See the Qwen model card for details.

Author

Dr. Sanjay Anbu — creator of Sentiance and the fused mind. Code & runtime: github.com/sanjaydoc/Sentiance.

Citation

@software{sentiance_fused,
  title  = {Sentiance: a functional cognitive architecture with a
            cognition-conditioned language model (the fused mind)},
  author = {Dr. Sanjay Anbu},
  year   = {2026},
  url    = {https://github.com/sanjaydoc/Sentiance}
}

CryptoGod97/sentiance-film

Model card — Sentiance fused mind (sentiance-fused)